I've been working on my talk for p99conf.io and because I wanted to re-use visuals from my ngrok.com/blog/quantiz... blog post, I used Remotion.
I didn't want to have to manually move clips around, though, because that's hard with eye tracking, so I made a DSL to switch slides based on what I say.
![A TypeScript codeblock that reads:
// Draft Deck.pptx — slides 28–31: Benchmark setup through 4-bit KL-divergence results.
{
camera: benchmarkThroughResults
.fromWords("so with that out of the way")
.toWords("the quantization process has actually done some damage here to the model")
.trimEnd(seconds(0.2)),
slides: [
{
component: slide.SingleTokenKLD,
until: words("so in practice when we're calculating the KL divergence").start(),
},
{
component: slide.WikitextBenchmark,
until: words("so here are the commands to do that").start(),
},
{
component: slide.createLlamaCppKLDCommands({
lineHighlights: [
{ at: words("install lama cpp").start(), lines: [1] },
{ at: words("download some model").start(), lines: [2] },
{ at: words("there is a llama dash quantize").start(), lines: [3] },
{ at: words("lama perplexity").start(), lines: [4] },
{ at: words("then you run lama perplexity again").start(), lines: [5] },
],
}),
until: words("this is the actual output").start(),
},
{
component: slide.FourBitKLDStatistics,
},
],
},](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:kcggckk55s6njoewaberea2l/bafkreidpivysoerzd2zo2k75mk22aszreapodx6dte5vzizeu6uk6lxcbi)
